With air quality being one target in the sustainable development goals set by the United Nations, accurate monitoring also of indoor air quality is more important than ever. Chemiresistive gas sensors are an inexpensive and promising solution for the monitoring of volatile organic compounds, which are of high concern indoors. To fully exploit the potential of these sensors, advanced operating modes, calibration, and data evaluation methods are required. This contribution outlines a systematic approach based on dynamic operation (temperature-cycled operation), randomized calibration (Latin hypercube sampling), and the use of advances in deep neural networks originally developed for natural language processing and computer vision, applying this approach to volatile organic compound measurements for indoor air quality monitoring. This paper discusses the pros and cons of deep neural networks for volatile organic compound monitoring in a laboratory environment by comparing the quantification accuracy of state-of-the-art data evaluation methods with a 10-layer deep convolutional neural network (TCOCNN). The overall performance of both methods was compared for complex gas mixtures with several volatile organic compounds, as well as interfering gases and changing ambient humidity in a comprehensive lab evaluation. Furthermore, both were tested under realistic conditions in the field with additional release tests of volatile organic compounds. The results obtained during field testing were compared with analytical measurements, namely the gold standard gas chromatography mass spectrometry analysis based on Tenax sampling, as well as two mobile systems, a gas chromatograph with photo-ionization detection for volatile organic compound monitoring and a gas chromatograph with a reducing compound photometer for the monitoring of hydrogen. The results showed that the TCOCNN outperforms state-of-the-art data evaluation methods, for example for critical pollutants such as formaldehyde, achieving an uncertainty of around 11 ppb even in complex mixtures, and offers a more robust volatile organic compound quantification in a laboratory environment, as well as in real ambient air for most targets.

In this study, methods from the field of deep learning are used to calibrate a metal oxide semiconductor (MOS) gas sensor in a complex environment in order to be able to predict a specific gas concentration. Specifically, we want to tackle the problem of long calibration times and the problem of transferring calibrations between sensors, which is a severe challenge for the widespread use of MOS gas sensor systems. Therefore, this contribution aims to significantly diminish those problems by applying transfer learning from the field of deep learning. Within the field of deep learning, transfer learning has become more and more popular. Nowadays, building a model (calibrating a sensor) based on pre-trained models instead of training from scratch is a standard routine. This allows the model to train with inherent information and reach a suitable solution much faster or more accurately. For predicting the gas concentration with a MOS gas sensor operated dynamically using temperature cycling, the calibration time can be significantly reduced for all nine target gases at the ppb level (seven volatile organic compounds plus carbon monoxide and hydrogen). It was possible to reduce the calibration time by up to 93% and still obtain root-mean-squared error (RMSE) values only double the best achieved RMSEs. In order to obtain the best possible transferability, different transfer methods and the influence of different transfer data sets for training were investigated. Finally, transfer learning based on neural networks is compared to a global calibration model based on feature extraction, selection, and regression to place the results in the context of already existing work.

Trust in artificial intelligence (AI) predictions is a crucial point for a widespread acceptance of new technologies, especially in sensitive areas like autonomous driving. The need for tools explaining AI for deep learning of images is thus eminent. Our proposed toolbox Neuroscope addresses this demand by offering state-of-the-art visualization algorithms for image classification and newly adapted methods for semantic segmentation of convolutional neural nets (CNNs). With its easy to use graphical user interface (GUI), it provides visualization on all layers of a CNN. Due to its open model-view-controller architecture, networks generated and trained with Keras and PyTorch are processable, with an interface allowing extension to additional frameworks. We demonstrate the explanation abilities provided by Neuroscope using the example of traffic scene analysis.

Data-driven analysis for damage assessment has a large potential in structural health monitoring (SHM) systems, where sensors are permanently attached to the structure, enabling continuous and frequent measurements. In this contribution, we propose a machine learning (ML) approach for automated damage detection, based on an ML toolbox for industrial condition monitoring. The toolbox combines multiple complementary algorithms for feature extraction and selection and automatically chooses the best combination of methods for the dataset at hand. Here, this toolbox is applied to a guided wave-based SHM dataset for varying temperatures and damage locations, which is freely available on the Open Guided Waves platform. A classification rate of 96.2% is achieved, demonstrating reliable and automated damage detection. Moreover, the ability of the ML model to identify a damaged structure at untrained damage locations and temperatures is demonstrated.

Selecting an appropriate model for industrial condition monitoring is challenging due to various factors. Typically, industrial datasets are small and lack statistical independence because experimental coverage of all possible operational variations is costly and sometimes practically impossible. Consequently, the resulting domain shifts pose a significant challenge. Although deep learning (DL) methods have frequently been regarded as the primary and optimal choice in many applications, they often lack major success factors in condition monitoring tasks. In this study, we benchmark the robustness of typical DL architectures against classical feature extraction and selection followed by classification (FESC) methods under domain shifts commonly encountered in industrial condition monitoring. Both DL and FESC methods are employed within an automated machine learning framework. We benchmarked these methods on seven publicly available datasets, and to simulate domain shifts, we employed leave-one-group-out validation on those datasets. Our experiments demonstrate high accuracy across all tested models for random K-fold cross-validation. However, the overall performance significantly decreases when faced with domain shifts, such as transferring the trained model from one machine to another. In four out of seven datasets, FESC methods showed better results in the presence of domain shifts. Furthermore, we also show that FESC methods are easier to interpret than DL methods. Finally, our results suggest that deep neural networks are not universally preferred over classical, low-capacity models for such tasks, as typically only a limited number of features from the input signal are needed.

Nowadays machine learning methods and data-driven models have been used widely in different fields including computer vision, biomedicine, and condition monitoring. However, these models show performance degradation when meeting real-life situations. Domain or dataset shift or out-of-distribution (OOD) prediction is mentioned as the reason for this problem. Especially in industrial condition monitoring, it is not clear when we should be concerned about domain shift and which methods are more robust against this problem. In this paper prediction results are compared for a conventional machine learning workflow based on feature extraction, selection, and classification/regression (FESC/R) and deep neural networks on two publicly available industrial datasets. We show that it is possible to visualize the possible shift in domain using feature extraction and principal component analysis. Also, experimental competition shows that the cross-domain validated results of FESC/R are comparable to the reported state-of-the-art methods. Finally, we show that the results for simple randomly selected validation sets do not correctly represent the model performance in real-world applications.